diff --git a/vllm/v1/attention/backend.py b/vllm/v1/attention/backend.py
index 28d077fcb771..c918958b855f 100644
--- a/vllm/v1/attention/backend.py
+++ b/vllm/v1/attention/backend.py
@@ -388,6 +388,12 @@ class CommonAttentionMetadata:
     (num_computed_tokens < num_prompt_tokens). Used by some backends to
     distinguish actual decodes from short extends."""
 
+    num_prompt_tokens: torch.Tensor | None = None
+    """(batch_size,) int tensor: original prompt length for each request.
+    Needed by dual-cache RoPE implementations (e.g. LongRoPE SplitByLength)
+    to select the correct cache per-sequence under chunked prefill, where
+    positions.max() of a chunk does not reflect the full prompt length."""
+
     # WARNING: Deprecated fields. Will be removed in a future release (v0.15.0)
     _seq_lens_cpu: torch.Tensor | None = None
     _num_computed_tokens_cpu: torch.Tensor | None = None
@@ -470,6 +476,7 @@ def unpadded(
             dcp_local_seq_lens=maybe_slice_reqs(self.dcp_local_seq_lens),
             dcp_local_seq_lens_cpu=maybe_slice_reqs(self.dcp_local_seq_lens_cpu),
             is_prefilling=maybe_slice_reqs(self.is_prefilling),
+            num_prompt_tokens=maybe_slice_reqs(self.num_prompt_tokens),
         )
 
 
diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index 4f3e192772ff..32bc189ba744 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -2192,6 +2192,7 @@ def _get_block_table(kv_cache_gid: int):
             slot_mapping=slot_mapping_gid_0,
             causal=True,
             is_prefilling=is_prefilling,
+            num_prompt_tokens=num_prompt_tokens_cpu,
         )
 
         if self.dcp_world_size > 1: